Code of the Bachelors Thesis: DEEP LEARNING STUDY FOR PULMONARY ILLNESS DIAGNOSIS

Table of Contents

1. Project Overview
2. Installation
3. Usage
   • Training
   • Evaluation
   • Prediction
4. Plotting Results
5. Configuration Flags
   • Main Script (main.py)
   • Plot Script (plot.py)
6. Results
7. License

---

Project Overview

This is the implementation of the code used for the Bachelors Thesis of Hugo Aranda at Universitat Politècnica de Catalunya. It includes an implementation of a particular Deep Learning architecture for the CheXRay-14 dataset presented in the paper by Wang et al. (2017) ChestX-Ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases.

The implementation has a variety of configuration options through flags that can be used in order to modify the Deep Learning architecture to fit various alternatives.

The associated documentation published on the UPC commons page is on the Memoria_HugoAranda_TFG.pdf document. This document contains all of the different experiments that generated the best performing model, the theoretical background for the project and different project management elements.

---

Installation

To get started, clone the repository and install the necessary dependencies using the provided environment.yml file.

git clone https://github.com/hugoarsa/TFG.git
cd TFG
conda env create -f environment.yml
conda activate cuda_dev_1

In order for the model to work you need to load the original datset images from the NIH website on the "/resized_images" folder

Also NVIDIA GPU drivers need to be up to date in order to use the GPU acceleration.

---

Usage

Training

The best performing model is included as a checkpoint but in order to retrain the model altogether you can use:

python main.py --mode train

This will use default parameters, you can adjust other training parameters using the available configuration flags.

Evaluation

To evaluate a trained model you can provide the model and prediction model save folder:

python main.py --mode eval --model efficientb0 --pred_model final_efficientb0_asl2_cyclic_SGD_60

Prediction

In order to generate predictions and GradCAM:

python main.py --mode pred --pred_model final_efficientb0_asl2_cyclic_SGD_60 --pred_image ./sample_images/00000211_004.png

---

Plotting Results

To visualize the history metrics from the training process you can use the plot.py script:

python plot.py --saves_dir ./saves --output_dir ./docu --target_var val_auc --name performance_plot.png

If you want to filter through the folders within saves you can use the --folders argument followed by a list of folders.

---

Configuration Flags

Main Script (main.py)

The following flags control the behavior of the model during training, evaluation, or prediction:

Flag	Default	Description	Choices
--mode	train	Mode of operation: train, evaluation, or prediction	['train', 'eval', 'pred']
--image_dir	./resized_images	Directory containing the input images	N/A
--pred_model	final_efficientb0_asl2_cyclic_SGD_60	Model checkpoint to use for predictions and evaluation	N/A
--pred_image	00000211_004.png	Image file for prediction	N/A
--train_csv	./labels/train_metadata.csv	Training metadata	N/A
--val_csv	./labels/val_metadata.csv	Validation metadata	N/A
--test_csv	./labels/test_metadata.csv	Test metadata	N/A
--save_dir	./saves	Directory where logs and model checkpoints will be saved	N/A
--model	efficientb0	Neural Network model to be used	['res18', 'res50', 'dense121', 'efficientb0', 'efficientb4']
--untrained	False	Specify whether the model is untrained (not pre-trained on ImageNet)	True/False
--thresh	False	Use dynamic threshold	N/A
--max_epochs	60	Maximum number of epochs for training	N/A
--num_iter	None	Maximum iterations per epoch	N/A
--batch_size	32	Batch size for data loaders	N/A
--img_size	256	Desired image size for the dataset	N/A
--loss	asl2	Loss function to use	['bce', 'bce_w', 'focal', 'asymmetric', 'asl1', 'asl2', 'asl3']
--lr	5e-4	Learning rate	N/A
--opt	Adam	Optimizer algorithm	['SGD', 'Adam', 'AdamW', 'RMSprop']
--scheduler	plateau1	Learning rate scheduler to use	['plateau', 'cyclic', 'cosine', 'warmupcosine']
--e_patience	15	Patience for early stopping	N/A
--s_patience	6	Patience for the learning rate scheduler	N/A

Plot Script (plot.py)

The following flags are available for generating plots:

Flag	Default	Description
--saves_dir	./saves	Directory where model checkpoints are saved
--output_dir	./docu	Directory where the plots will be saved
--target_var	val_auc	Variable to plot (e.g., validation accuracy or loss)
--name	loss_selection_plot.png	Name of the output plot file
--folders	None	List of specific folder names to include in the plot

---

Results

See results at the published Memory for the project in the upc commons webpage.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.